Skip to content

Improve recall reliability and introduce the update on the homepage - #14

Merged
focaxisdev merged 6 commits into
mainfrom
fix/recall-efficiency-and-budgets
Oct 18, 2026
Merged

focaxisdev merged 6 commits into
mainfrom
fix/recall-efficiency-and-budgets

Conversation

@focaxisdev

Copy link
Copy Markdown
Owner

Eligible memories could be hidden by ineligible hybrid-ranking winners, while repeated lexical preprocessing and unbounded prompt payloads added avoidable cost. This change makes eligibility precede hybrid preference, separates candidate breadth from returned output, validates limits, and adds bounded prompt projection without changing the three-file protocol or default familiarity thresholds.

Optional aliases, metadata-only embedding reuse, bounded content/model caches and managed ownership add explicit control over routing and cache lifetime. Update/delete/supersede recovery and projection provenance are covered by regression tests. README now presents the source update prominently, preserves the 60-second adoption path, and links opt-in/migration guidance and scoped performance/quality evidence. A new npm package or GitHub Release is not part of this PR.

Validation:

  • Prior exact-tree branch run 39089843480 passed on Ubuntu and Windows: 119 source, 119 compiled and one package/install/use case per platform, zero skipped; types, frozen quality/cache counts and documentation also passed.
  • This head adds the homepage/guide and reuses the same full acceptance workflow for main/PR, including fresh synthetic-project lint/readiness. Its branch and PR checks must pass before merge.
  • All original local commits remain preserved. Atomic connector publication maps local 30d32751613557d7b3d252040b07aa23317b2171 to this remote head; complete tracked tree 4147bece082ecb5a44ae48c21aee9da852999bbe was verified path/type/mode/blob by path.

Measurement limits remain explicit: 30.74 → 2.61 ms is one 10,000-record common-cue search microbenchmark; 2/13 → 10/13 is a small same-author synthetic alias evaluation. The second batch's roughly 12.4% resident-API latency increase was measured in default mode, not an isolated ownership-mode benchmark. No global speedup, multilingual semantic model or secure-erasure claim is made.

No force push, dependency additions, permission changes, deployment or package publication.

Atomic connector publication of the exact tracked tree from local commit a66a52d4fceaeba3dac4a24c8e7cc82d834a4331.
Source tree: 7365590
Verified remote main parent: 188d7e8
Original local commits bad6b527409a44a693562fbe60c41be74126e573, fa479335dbc5acbad46a352d1ee8bd6086c62cd7, 4c05a77605f6cf63a10261dc394c0824208313ea and later checkpoint commits remain preserved locally. This new remote commit intentionally has a different SHA.

No release, merge or package publication. Branch-only acceptance must pass independently.
Exact source local commit: f880ca2d67b14e939164bb7fc594800598e8fcdd
Verified tracked tree: da94ba9
Original local history retained; remote SHA intentionally differs.
…ort paths

Exact source local commit: bb2b64e486e8530c8eeca09ed3d8fe6fc71f8c67
Verified tracked tree: c574e26
Original local history retained; remote SHA intentionally differs.
Exact source local commit: b139302c87b891cdc4e45ed735b5e2b7461b941f
Verified tracked tree: 0cbc219
Original local history retained; remote SHA intentionally differs.
Exact source local commit: 7ccb05ceb29cd67940dfee64f182a875f1326a9d
Verified tracked tree: ea4f5e4
Original local history retained; remote SHA intentionally differs.
Exact source local commit: 30d32751613557d7b3d252040b07aa23317b2171
Verified tracked tree: 4147bec
Original local history retained; remote SHA intentionally differs.
@focaxisdev
focaxisdev marked this pull request as ready for review October 18, 2026 02:38
@focaxisdev
focaxisdev merged commit b8ff7d5 into main Oct 18, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant